The system was put in charge of two logical qubits hosted on a calibrated system. The two were using different error correction schemes (a surface code and a color code). These were set in a specific state, and the error-correction system was then used with and without reinforcement-learning-driven corrections. Having the system active led to a 20 percent increase in the ability to detect and correct errors in the logical qubits.
Going real time
The limitation of this approach is that it works only if the drift keeps the system reasonably close to the state the system was trained in. The corrections that might bring things back into alignment from one state might not be effective when the system’s in a significantly different state.
The solution to this is to constantly re-evaluate the effectiveness of different changes. But this has an obvious problem: You can’t simply randomize all the potential control configurations in the middle of a calculation. Even with limited variation, the system will necessarily operate outside its optimal error correction. So, the question was whether the frequent sub-optimal error correction paid off by keeping drift from causing even larger problems. “The favourable resolution of the exploration–exploitation trade-off would mean that the aggregate performance of all sampled policy candidates, most of which are worse than [the optimal one], is still better than the performance without reinforcement learning steering,” the researchers write.
Performing many simulations with a very small error-corrected qubit showed that the trade-off worked out, provided that drift was slow enough. The team showed that it could work in real time with a large error-corrected qubit, in which the reinforcement learning system had control over roughly 40,000 parameters.
This is clearly not a solution for the present; we can only keep systems operating for long enough to perform relatively short, simple algorithms, so drift isn’t even a concern. Ultimately, our intention is to build hardware that can perform the sorts of calculations where issues like this will matter. And there’s some value in demonstrating that something we know could be a problem can be dealt with.
Nature, 2026. DOI: 10.1038/s41586-026-10759-2 (About DOIs).
